Impact of Popular Content Relational Structure on Joint Caching and Recommendation Policies

Recent work has shown that the performance of caching systems can be boosted by the delivery of alternative but related content. In this setup, recommendation systems are exploited to offer the user appealing alternatives if the original request is not found in the cache. This framework relies on the assumption that a content can partially or completely substitute another if they are sufficiently similar. In this work we model these similarity relations as a graph, where each content is a node and related contents are linked by an edge. We then study how the characteristics of this content graph constrain the gains of designing jointly the caching and the recommendations with respect to just using a simple baseline in the soft cache hits setup. We start by selecting a number of descriptive graph features that we expect to play a role in the performance of related content delivery policies. We then analyze the effect of each of these features on the policies' performance through extensive simulations with synthetic data. Our results confirm that features such as the degree distribution and clustering coefficient of the graph are crucial to decide whether the close-to-optimal algorithm will perform significantly better than the simple baseline. Our experiments with four real-world distinct datasets further support these observations. Motivated by these clear dependencies, we conclude by showing that we can train a classifier to predict the gains attainable by a “smart” policy with respect to a low-complexity baseline using only a few macroscopic graph features as predictor variables.